211–220 of 236 posts
I think the real problem is most people don't actually have a very good understanding of "Truth." As someone who ended up studying philosophy, there seems to be a real gulf between folks who sort of believe stuff they hear, folks who believe "facts" that they hear from (various levels of) credible sources, and folks that take solipsism seriously understand that even in the most ideal scenario, we still wouldn't have…
Earlier quoted context omitted.
> Is Google search engine that leads to NY Times or Fox News or Wikipedia and makes us manually choose sources as per our biases "better" than Google's Gemini engine that summarizes content from all the above sources and gives an average answer? If you use just any amount of critical thinking, yes. Truth and objectivity are ideals, not practical states. LLMs are a very bad way to come close to this ideal. You may use…
> LLMs are a very bad way to come close to this ideal...the output directly is a strict degeneration I didn't understand the second part but regarding the first... For me, LLMs are just another source of information with a different UI, analogous to newspapers, TV documentaries, Wikipedia, Google search, YT talks/documentaries, even the majority of informational non-fiction books, and research papers. Some may consid…
With all due respect (not trying to be offensive at all) but this is insane to me.
All those sources of information you cited have a million incentives to provide fairly correct and checked information. But probably more importantly, they have even more incentives to NOT provide false information. At a minimum, their careers, reputation, recurring work, brand, etc... is on the line.
An LLM has zero incentives to provide you with true information, beyond a couple of md files with instructions. If it gets it wrong, there is zero accountability, just an -oh, you're absolutely right- response and move on.
I agree there is a lot of human bias in the world, but surely we can't even put in the same order of magnitude both types of biases!
My method is simple. I remind them that chatgpt is trained on everything said on the internet including NYT if speaking to a Republican, replace that with Fox News if speaking to a Democrat.
I like a lot of the other answers, but this is clearly the most pragmatic one.
Earlier quoted context omitted.
Even as a human, I find whatever sources Google shows to be inconsistent. I can't give any confident answer about the number of fingers. I think the answer is actually "4 sometimes and 5 other times." So I'm not sure how much LLMs can handle this kind of inconsistency between "reputable" visual sources and text sources, nor how representative this example is. A "reputable source" like Variety says this... https://var…
Your first link was cited in the 2nd half of Google AI's answer, and one of the top Google answers, so I think you are using Google as your information source. The large majority of the images you link to show kids with 5 fingers, as well as 5-fingered baseball gloves. The cases of four fingers are due to orientation. Your "1." also shows Marcie with five fingers. You see Charlie Brown with 4 fingers because he's hol…
Earlier quoted context omitted.
While some of the ideas in this do resonate with me (or at least they're entertaining), it's unfortunate that's it's so obviously LLM generated. And some parts of it, like the INTJ exceptionalism, reek of LLM sycophancy, which then turned into to some kind of god complex...
observation a: Document title is about a minority's rightful supremacy observation b: document says "this is not political" then dives into persuasive speech conclusion: this document was written by the bad guys
Earlier quoted context omitted.
Claude Opus 4.6 is the best possible model to use in this test, with the least sycophancy. OpenAI and Gemini models are bad in comparison.
ChatGPT thinking models are very good; the instant model is bad. Gemini is always desperate to find an answer, and will give you one no matter what.
Earlier quoted context omitted.
ChatGPT thinking models are very good; the instant model is bad. Gemini is always desperate to find an answer, and will give you one no matter what.
Nope, I use GitHub Copilot (agentic mode) and I end up having to use the (more expensive) Claude model because ChatGPT never second-guesses me or even itself. Gemini is slightly worse though.
Earlier quoted context omitted.
Well, in my case going to a Peanuts comic and looking at hands was pretty easy, and didn't involve any questions about negative environment or labor consequences, the massive hammering of web sites to gather data, centralization of power, and the like. Like, "!w Peanuts" in my search bar, look at the image, and count fingers. "a rather outlier example" You wrote that you use AI to find "obscure connections" - aren't…
HN isn't showing me a reply option for your latest comment, so I'll reply here instead. Just to clarify, I used plain Google search not Google AI mode. And opened search results which seemed "reputable," without knowing anything much about Peanuts cartoon or cartooning. I had no idea at all about archive.org having it and didn't see it listed in the first two pages of search results. I still find it confusing, especi…
I think the right way to interpret the Variety link is that it's a single paragraph about trying to capture the feel of the comic using 3D software. As you saw from Charlie Brown holding a baseball, Shulz didn't go for a realistic look, but still conveys the sense of grasping. Modeling all five fingers all the time would not give the movie the right feel.
I wonder now if Google AI incorporates text from the top results into its answer.
"why is it wrong when the AI generated 4 fingers?"
The original discussion was when person X used AI to generate a image "in the style f Charles Shulz" where the Peanuts characters had 5 fingers, then person Y noted the use of 5 fingers instead of the 4 which is common in comics and cartoon, and quoted Google AI as saying Peanuts was traditionally drawn with 4 fingers.
I yesterday verified that Google AI would generate the same wrong answer with a text query, so it was not an image interpretation issue.
FWIW, after looking at a few hundred Peanuts cartoons, I can confidently say the AI generated image was not in the style of Schulz. The generated fingers were too realistic, and the background too complicated and detailed. :)
This for me is another example of why using primary sources should be the first thing to consider when fact checking - not LLMs (my experience is they are horrible at details), and not secondary sources (which have their own biases).
Not everything has easily-accessed primary sources, but many do. I think it's all too easy to fall into the trap of accepting the LLM answer because it feels right and is easy to generate. At https://freethoughtblogs.com/stderr/2025/01/18/ai-art-just-r... you'll see someone asked about which river Marbot swam across to spy on the enemy camp. It replied "Elbe". Then I did a text search of an English translation of the book and found he used a boat to cross the Danube to spy on the enemy camp, and he swam into freezing waters to save an enemy soldier.
Again, do you ask the LLM to fact check itself every single time? If that's useful, why isn't it built into the prompt? Or, if you are supposed to double-check the LLM yourself, why would you consult a secondary source if the primary source is so easy to find and search? And in that case, why not just use the primary source?
Further, if you aren't in the habit of checking primary sources then you won't have the experience to know how to find and check primary sources.